fix(extract): normalize whitespace before truncating rationale labels (#2206)

[Yyunozor]

Summary

_extract_python_rationale / _extract_js_rationale build rationale node labels by slicing the raw docstring/comment text to 80 characters before normalizing whitespace. This can cut a word in half, leave literal spaces where a newline + indentation used to be, and — when the cut lands on a . — produce an Obsidian filename ending in ..md. Not export-only: it's the label in graph.json, so the artifacts show up in GRAPH_REPORT.md, graphify query, and explain too.

Reproduction

def reconcile_batches():
    """This routine reconciles pending settlement batches nightly because upstream feeds arrive unordered and out of sequence."""

Before: label ends "...because upstream feed" (80 chars, "feeds" sliced mid-word). A docstring whose 80th char lands on a sentence-ending . exports as ...and caches..md.

Root cause

Both _add_rationale closures in graphify/extract.py do:

label = text[:80].replace("\r\n", " ").replace("\r", " ").replace("\n", " ").strip()

The slice runs before whitespace normalization, so it can land inside a word or a newline's collapsed space run, and .strip() only trims surrounding whitespace, never a trailing period.

Fix

Both sites now call a shared _shorten_rationale_label:

label = textwrap.shorten(text, width=width, placeholder="…")
if label in ("", "…"):
    flat = " ".join(text.split())
    label = flat if len(flat) <= width else flat[: width - 1] + "…"

textwrap.shorten collapses whitespace, then drops whole words from the end (never mid-word), adding the placeholder only when it truncates — a short label is unchanged. The fallback guards a shorten quirk: a first word alone >= 80 chars (e.g. one long URL opening a comment) collapses to a bare "…", worse than the parent's raw slice. The fallback restores content.

Known limitation

shorten leaves an untruncated label's own trailing punctuation alone, so a short rationale legitimately ending in . still hits a separate, pre-existing gap: export.py's two safe_name() (~L500, ~L801) strip .md/.mdx/.qmd/.markdown but not a bare .. Different file, out of scope here — a follow-up would strip a trailing . in safe_name, as it already strips a leading one (#2205).

Validation

Nine new tests in tests/test_rationale.py, split across Python and JS/TS: word-boundary truncation, whitespace-run collapse, no-trailing-period, long-token fallback, short-docstring non-regression. All fail on parent with the reported symptom, pass after the fix.

python -m pytest tests/ -q

Delta is exactly these 9 tests; skip set identical both sides. Cross-checked on Python 3.10. ruff check and all five tools.skillgen guards clean; graphify update smoke run OK.